Identification of disease using deep learning and evaluation of bacteriosis in peach leaf

• Published in: Ecological informatics 2021-03, Vol.61, p.101247, Article 101247
• Main Authors: Yadav, Saumya, Sengar, Neha, Singh, Akriti, Singh, Anushikha, Dutta, Malay Kishore
• Format: Article
• Language: English
• Subjects: Bacterial spot; Deep learning; Image processing; Morphological processing; Peach crops
• ISSN: 1574-9541
• DOI: 10.1016/j.ecoinf.2021.101247

Bacteriosis is one of the most common and devastating diseases for peach crops all over the world. Timely identification of bacteriosis disease is necessary for reducing the usage of pesticides and minimize loss of crops. In this proposed work, convolutional neural network (CNN) models using deep learning and an imaging method is developed for bacteriosis detection from the peach leaf images. In the imaging method, disease affected area is quantified and an adaptive operation is applied to a selected suitable channel of the color image. Gray level slicing is done on pre-processed leaf images for segmentation and automatic identification of bacterial spot disease in peach crops. The datasets are augmented to make the algorithm more robust to different illumination conditions. The proposed work compares the result of imaging method and CNN method. Model architectures generated with different deep learning algorithms, had the best performance reaching an accuracy of 98.75%% identifying the corresponding peach leaf [bacterial and healthy] in 0.185 s per image. The test dataset is consist of images from real cultivation field and also from the laboratory conditions. The significantly high identification rate makes the model diagnostic or early warning tool, and an approach that could be further integrated with the unmanned aerial vehicle to operate in real farming conditions •Deep learning model and imaging method is proposed for identification of bacteriosis form peach leaf.•An imaging method developed to quantify the disease affected area from the leaf.•Automatic segmentation of the disease affected area by image processing method.•The final model achieved 98.75% accuracy on 240 previously “unseen” images from lab and real cultivation field.